/sci/ - Science & Math » Thread #11657798

77KiB, 2400x1000, openai-1.png

View Same Google iqdb SauceNAO

What's the point of trying to do AI work when people with billions in compute will outdo you?

Anonymous Sun 10 May 22:48:34 2020 No.11657798 View Reply Original Report

Quoted By: >>11657876 >>11657924 >>11659596 >>11661829 >>11662011 >>11662792 >>11663179 >>11666017 >>11667305

Anonymous

Anonymous Sun 10 May 2020 23:07:08 No.11657876 Report

Quoted By: >>11657919

>>11657798
there are tasks that even infinite linear compute cant solve

Anonymous

Anonymous Sun 10 May 2020 23:20:17 No.11657919 Report

Quoted By:

>>11657876
Non responsive

Anonymous

Anonymous Sun 10 May 2020 23:22:12 No.11657924 Report

Quoted By:

>>11657798
>outdo you
why are you directly competing with giants?

Anonymous

View Same Google iqdb SauceNAO kl.jpg, 40KiB, 401x600

Anonymous Mon 11 May 2020 12:52:39 No.11659596 Report

Quoted By:

>>11657798
I am just a random guy living in northen Europe and I will make something that beats a proper Turing test before OpenAI, probably before the end of 2021.

Every sperg on 4chan will finally have his own AI gf.
Screencap this.

Anonymous

Anonymous Mon 11 May 2020 21:09:38 No.11660927 Report

Quoted By: >>11661373

Whats the point of trying to solve the problem od transmitting images when companies with billions in resources will outdo you?

"Idk they sound like fags" -Philo Farnsworth

Anonymous

Anonymous Mon 11 May 2020 23:32:04 No.11661373 Report

Quoted By:

>>11660927
>-Philo Farnsworth
who?

Anonymous

Anonymous Tue 12 May 2020 02:19:17 No.11661829 Report

Quoted By: >>11665244

>>11657798
Machine learning is starting to hit a lot of walls already. You could add an entire building of computer power to a Tesla and it wouldn't improve at all at this point.

Anonymous

Anonymous Tue 12 May 2020 03:19:23 No.11662011 Report

Quoted By: >>11662015 >>11662404 >>11662415 >>11662425

>>11657798
...do you think AI work is about who has the biggest computer?

Anonymous

Anonymous Tue 12 May 2020 03:20:01 No.11662015 Report

Quoted By: >>11662415

>>11662011
/thread

Anonymous

Anonymous Tue 12 May 2020 05:13:44 No.11662404 Report

Quoted By: >>11663733 >>11666564

>>11662011
>...do you think AI work is about who has the biggest computer?
Not OP, but yes.
You are ignorant of the industry and the state-of-the-art in research if you seriously believe otherwise.

Anonymous

Anonymous Tue 12 May 2020 05:16:11 No.11662415 Report

Quoted By:

>>11662011
>>11662015
imagine being this dumb

Anonymous

Anonymous Tue 12 May 2020 05:18:44 No.11662425 Report

Quoted By: >>11663733 >>11666570

>>11662011
ML is so strongly dependent on bruteforce that you could easily discard a bunch of the newest algorithms if you could have a bunch of GPU cycles at your free disposal.

Anonymous

View Same Google iqdb SauceNAO 1585360957573.png, 272KiB, 622x461

Anonymous Tue 12 May 2020 07:43:02 No.11662792 Report

Quoted By:

>>11657798
When are these CS dorks gonna program me an AI gf?

Anonymous

Anonymous Tue 12 May 2020 10:34:34 No.11663179 Report

Quoted By:

>>11657798
>outdo you
Then oversell them some novel approach you developed and get paid for consulting since they are hoarding the tech and have that much money to throw around.

Anonymous

Anonymous Tue 12 May 2020 14:49:22 No.11663733 Report

Quoted By: >>11665179 >>11665273

>>11662404
>>11662425
You still have to have good ideas. Having more computing power just means you can try out more things in less time. You can't just brute force everything because the search space is way too big.
At this point new fundamental theoretical insights would be much more useful than any experimental ones. But of course they are way harder to achieve.

Anonymous

Anonymous Tue 12 May 2020 21:51:41 No.11665179 Report

Quoted By: >>11666564 >>11666902

>>11663733
Ilya Sutskever actually showed how their NLP models started to grasp higher level things like semantics or meaning or sentiment only after huge number of parameters was reached

Also look into the bitter lesson from Richard Socher

Anonymous

Anonymous Tue 12 May 2020 22:08:22 No.11665244 Report

Quoted By: >>11665271

If you develop something worth running at scale, you can get compute for free from nvidia.

>>11661829
agree about tesla... full self driving is a total meme right now (LOL waymo just got 3bn more today... such a scam). Disagree that ML has hit a wall... RL and meta-learning is getting real interesting right about now

Anonymous

Anonymous Tue 12 May 2020 22:19:49 No.11665271 Report

Quoted By:

>>11665244
>If you develop something worth running at scale, you can get compute for free from nvidia.
Elaborate please
Do they also get to keep my code?

Anonymous

Anonymous Tue 12 May 2020 22:20:10 No.11665273 Report

Quoted By:

>>11663733
Agree bigly. In reality, a human baby is exposed to what, a few billion "frames" of experience (I know, I know, not an accurate comparison) before it can talk, walk and is self-aware. It's not the compute holding us back, it's the algorithms.

Anonymous

Anonymous Wed 13 May 2020 02:51:21 No.11666017 Report

Quoted By:

>>11657798
Think about your own brain.
The theory is we developed a capacity for language about 50,000 years ago.
That's how long it took for evolution to refine our brains
I'm pretty sure OpenAI won't discover anything except really nice statistical learning theory for certain niche things, like AR with your phone as a star trek 'tricorder' or something or absolutely horrible weapons/drones otherwise no actual AI

Anonymous

Anonymous Wed 13 May 2020 06:25:48 No.11666564 Report

Quoted By: >>11666566

>>11662404
>>11665179
The bitter lesson is not that more computing power = progress. It's that instead of you, as the researcher, not stroking yourself off by giving NLP and learning agents seemingly sophisticated strategies. The bitter lesson is that search and learn are sophisticated enough to scale well and pick up strategy, so you should focus on being mathematically elegant and precise on the meta-methods (ie, the sophisticated ways to perform learn and search). This way, you effectively scale the efficacy of your research as computational power improves.

The takeaway is to stop drinking dumb bitch juice and accept that search and learn are powerful. Stop jerking yourself off and thinking your single strategy is better than circuits going brr

Anonymous

Anonymous Wed 13 May 2020 06:26:50 No.11666566 Report

Quoted By:

>>11666564
*should not be stroking yourself off

Anonymous

Anonymous Wed 13 May 2020 06:30:01 No.11666570 Report

Quoted By:

>>11662425
No. ML is dependent on computational power, but applied ML research is chiefly in two areas - applying known techniques to new problems, and new methods to perform searching in a more sophisticated manner. The idea is that the computer can learn provided it can search, and searching and converging to an answer is studied in the PAC model / statistical learning theory. There's a lot of active research in methods in searching better - if you think ML research is reduced down to running libraries on GPUs, you've bought into the memes about ML research my friend

Anonymous

Anonymous Wed 13 May 2020 09:44:54 No.11666902 Report

Quoted By: >>11667134

>>11665179
>Ilya Sutskever
But he's literally one of the guys in charge of OpenAI so surely what you said proves that you don't need massive computing power to make significant contributions to AI development

Anonymous

Anonymous Wed 13 May 2020 12:03:19 No.11667134 Report

Quoted By:

>>11666902
What? Dude OpenAI is planning to scale GPT-2 by a factor of up to 1000x (both in data and parameters number)

Anonymous

Anonymous Wed 13 May 2020 13:26:55 No.11667305 Report

Quoted By:

>>11657798
If you find a method that is a decent factor quicker than the current AI, you won't need a giant ass computer. The limitations will force you to innovate.

Capcode	All Only User Posts Only Moderator Posts Only Admin Posts Only Developer Posts
Show Posts	All Only With Images Only Without Images
Deleted Posts	All Only Deleted Posts Only Non-Deleted Posts
Ghost Posts	All Only Ghost Posts Only Non-Ghost Posts
Post Type	All Only Sticky Threads Only Opening Posts Only Reply Posts
Results	All Grouped By Threads
Order	Latest Posts First Oldest Posts First

Your latest searches